The reckoning
One schema change (captures stop being proto-tasks) and the five modes fall out of it.
On the 10th of May 2026, the pivot the whole project had been walking toward finally landed, in one document drop (the v3 spec plus six architecture decision records, in a single commit).
What it was. A change that sounds small and changes everything: a captured thought stopped being a task-in-waiting. Captures became their own first-class thing, and a task became just one possible outcome, something that only appears if the right mode produces one (captures ≠ tasks, separated at the schema level; in the ADR’s words, “the substrate change that makes everything downstream cohesive”).
Everything Alter is now fell out of that decision within days:
- Recognition: instead of one funnel, the system first works out what kind of thought this is (a fast classification that answers in under a second, with deeper enrichment continuing in the background), then routes it to one of five surfaces, each built for its kind: file it, answer it, think it through, build it out, or research it.
- The wiki as the durable layer: the workbench gave way to a self-compiling wiki: the stable place your knowledge accumulates, quietly maintained for you.
- Foundations, not features: encryption of your content by default, and a privacy-guarded event log, went in at the base layer where they can’t be forgotten.
Where the wiki idea came from: Andrej Karpathy’s notion of an LLM quietly keeping a clean, self-compiling knowledge base out of a messy stream of inputs, which is exactly the job Alter’s wiki does.
Why it took three products to learn. chotu stored blobs. CODE forced tasks. Only classifying intent first finally matched the shape of the response to the shape of the thought. Some lessons apparently need to be paid for three times.
Next era: Greenfield: 72 hours